Safety Switch

Written by Brook Schaaf

Last week may have marked a critical shift in how businesses view AI. In a widely-referenced CNBC interview, Alex Karp, the high-energy CEO of Palantir, argued that businesses are “unhappy with the frontier labs” (e.g., Anthropic, OpenAI) because of a “level of discomfort and distrust” over pricing and IP sharing. What customers want instead is control over their compute, costs, and intellectual property, which he just so happens to offer through Palantir’s new Nvidia partnership. (Duly acknowledged as “slightly self-centered.”)

Unfavorable reporting headlines ranged from “Heated Interview” to “Televised Nervous Breakdown.” In contrast, the supportive hosts of the All-In Podcast argued that he is taking flak because he’s over the target: even if still unprofitable, frontier lab costs are excessive compared to near-par, open-source models. Indeed, more importantly, the enterprise AI providers threaten to become direct competitors, as with Anthropic launching Claude Design. Karp stated his customers are buying the idea of his new offering “because these things have to be made safe.”

Does that last word sound familiar? It should. The word has been central to Anthropic’s identity ever since “Safety Crusader” Dario Amodei left OpenAI in 2020 after internal disagreements. 

He is quoted to have said, “When you feel that you can’t trust someone, when you feel that their values are not what they say they are, when you feel that they’re not honest, that makes it very hard to continue to work with a company.”

He may now be hoisted by his petard, if the definition of “safety” has again expanded. When I was young, it generally referred to a protocol for or expectation of reduced physical danger. By and by, the word came to refer to one’s emotional comfort level, then the protection of information and communications (think “brand safety” or misinformation). Here, Karp and his friends have tacked on the security of intellectual property and mitigation of competition to the term. Thus, he can use the same word Anthropic made so much headway with, even if it’s in a different sense. Moreover, he can argue that any AI provider with this business model is unsafe for business because it exposes critical IP and invites competition, like selling a popular product on Amazon only to watch Amazon produce its own version.

So first ChatGPT, then Gemini, now Claude, but perhaps tomorrow, the fastest horse will be open-source. And, who knows? Perhaps “safety” will eventually come to include safeguarding the publishers whose intellectual property powers these models.